Codegen for invoke-super, invoke-interface
Completed fast & slow paths for invoke-super and the single path for invoke-interface. Added test
for invoke-super, but invoke-interface is untested (needs runtime-support routines to be fleshed
out).
Some trickiness in the invoke-interface code-generation. Because they are going to be
glacially slow anyway, inline code has been minimized and all interesting work pushed
off to TODO runtime support routines. However, we can't simultaneously pass the arguments
needed by the final destination and the arguments needed by the runtime lookup helpers. So,
I've added a trampoline to save the target args, load the args needed by the helpers, call
the lookup routines, restore the final target arguments and continue on the journey.
More detailed comments in the code.
Change-Id: Ice2343798a91a37da982811fd1c6384f584a3c0b
diff --git a/src/compiler/codegen/arm/MethodCodegenDriver.cc b/src/compiler/codegen/arm/MethodCodegenDriver.cc
index d95b870..eb9296b 100644
--- a/src/compiler/codegen/arm/MethodCodegenDriver.cc
+++ b/src/compiler/codegen/arm/MethodCodegenDriver.cc
@@ -478,7 +478,7 @@
skipTarget = newLIR0(cUnit, kArmPseudoTargetLabel);
skipTarget->defMask = ENCODE_ALL;
skipBranch->generic.target = (LIR*)skipTarget;
- // Get base_method->method_index [usr rLR, set r12]
+ // Get base_method->method_index [usr rLR, set r0]
loadBaseDisp(cUnit, mir, rLR,
Method::GetMethodIndexOffset().Int32Value(), r0,
kUnsignedHalf, INVALID_SREG);
@@ -528,62 +528,25 @@
return callState;
}
-/*
- * Interleave launch code for INVOKE_INTERFACE. The target is
- * identified using artFindInterfaceMethodInCache(class, ref, method, dex)
- * Note that we'll have to reload "this" following the helper call.
- *
- * FIXME: do we need to have artFindInterfaceMethodInCache return
- * a NULL if not found so we can throw exception here? Otherwise,
- * may need to pass some additional info to allow the helper function
- * to throw on its own.
- */
+// Interleave launch code for INVOKE_INTERFACE.
static int nextInterfaceCallInsn(CompilationUnit* cUnit, MIR* mir,
DecodedInstruction* dInsn, int state,
ArmLIR* rollback)
{
- UNIMPLEMENTED(FATAL) << "Need findInterfaceMethodInCache";
-#if 0
- RegLocation rlArg;
switch(state) {
- case 0:
- // Load "this" [set r12]
- rlArg = oatGetSrc(cUnit, mir, 0);
- loadValueDirectFixed(cUnit, rlArg, r12);
- // Get the current Method* [set arg2]
- loadCurrMethodDirect(cUnit, r2);
- // Is "this" null? [use r12]
- genNullCheck(cUnit, oatSSASrc(mir,0), r12,
- mir->offset, NULL);
- // Get curMethod->clazz [set arg3]
- loadBaseDisp(cUnit, mir, r2, OFFSETOF_MEMBER(Method, clazz),
- r3, kWord, INVALID_SREG);
- // Load this->class [usr r12, set arg0]
- loadBaseDisp(cUnit, mir, r12, OFFSETOF_MEMBER(Class, clazz),
- r3, kWord, INVALID_SREG);
- // Load address of helper function
- loadBaseDisp(cUnit, mir, rSELF,
- OFFSETOF_MEMBER(Thread, pArtFindInterfaceMethodInCache),
- rLR, kWord, INVALID_SREG);
- // Get dvmDex
- loadBaseDisp(cUnit, mir, r3, OFFSETOF_MEMBER(Class, pDvmDex),
- r3, kWord, INVALID_SREG);
- // Load ref [set arg1]
- loadConstant(cUnit, r1, dInsn->vB);
- // Call out to helper, target Method returned in ret0
- newLIR1(cUnit, kThumbBlxR, rLR);
+ case 0: // Load trampoline target
+ loadWordDisp(cUnit, rSELF,
+ OFFSETOF_MEMBER(Thread, pInvokeInterfaceTrampoline),
+ rLR);
+ // Load r0 with method index
+ loadConstant(cUnit, r0, dInsn->vB);
break;
- case 1: // Get the target compiled code address [use r0, set rLR]
- loadBaseDisp(cUnit, mir, r0, OFFSETOF_MEMBER(Method, compiledInsns),
- rLR, kWord, INVALID_SREG);
default:
return -1;
}
-#endif
return state + 1;
}
-
/*
* Interleave launch code for INVOKE_SUPER. See comments
* for nextVCallIns.
@@ -592,49 +555,132 @@
DecodedInstruction* dInsn, int state,
ArmLIR* rollback)
{
- UNIMPLEMENTED(FATAL) << "Need INVOKE_SUPER implementation";
-#if 0
+ DCHECK(rollback == NULL);
RegLocation rlArg;
+ /*
+ * This is the fast path in which the target virtual method is
+ * fully resolved at compile time. Note also that this path assumes
+ * that the check to verify that the target method index falls
+ * within the size of the super's vtable has been done at compile-time.
+ */
+ Method* baseMethod = cUnit->method->GetDexCacheResolvedMethods()->
+ Get(dInsn->vB);
+ CHECK(baseMethod != NULL);
+ Class* superClass = cUnit->method->GetDeclaringClass()->GetSuperClass();
+ CHECK(superClass != NULL);
+ int32_t target_idx = baseMethod->GetMethodIndex();
+ CHECK(superClass->GetVTable()->GetLength() > target_idx);
+ Method* targetMethod = superClass->GetVTable()->Get(target_idx);
+ CHECK(targetMethod != NULL);
switch(state) {
- case 0:
- // Get the current Method* [set r0]
+ case 0: // Get current Method* [set r0]
loadCurrMethodDirect(cUnit, r0);
// Load "this" [set r1]
rlArg = oatGetSrc(cUnit, mir, 0);
loadValueDirectFixed(cUnit, rlArg, r1);
- // Get method->clazz [use r0, set r12]
- loadBaseDisp(cUnit, mir, r0, OFFSETOF_MEMBER(Method, clazz),
- r12, kWord, INVALID_SREG);
- // Get pResmethods [use r0, set rLR]
- loadBaseDisp(cUnit, mir, r0, OFFSETOF_MEMBER(Method, pResMethods),
- rLR, kWord, INVALID_SREG);
- // Get clazz->super [use r12, set r12]
- loadBaseDisp(cUnit, mir, r12, OFFSETOF_MEMBER(Class, super),
- r12, kWord, INVALID_SREG);
- // Get base method [use rLR, set r0]
- loadBaseDisp(cUnit, mir, rLR, dInsn->vB * 4, r0,
- kWord, INVALID_SREG);
+ // Get method->declaring_class_ [use r0, set rLR]
+ loadWordDisp(cUnit, r0, Method::DeclaringClassOffset().Int32Value(),
+ rLR);
// Is "this" null? [use r1]
genNullCheck(cUnit, oatSSASrc(mir,0), r1,
mir->offset, NULL);
- // Get methodIndex [use r0, set rLR]
- loadBaseDisp(cUnit, mir, r0, OFFSETOF_MEMBER(Method, methodIndex),
- rLR, kUnsignedHalf, INVALID_SREG);
- // Get vtableCount [use r12, set r0]
- loadBaseDisp(cUnit, mir, r12,
- OFFSETOF_MEMBER(Class, vtableCount),
- r0, kWord, INVALID_SREG);
- // Compare method index w/ vtable count [use r12, use rLR]
- genRegRegCheck(cUnit, kArmCondGe, rLR, r0, mir->offset, NULL);
- // get target Method* [use rLR, use r12, set r0]
- loadBaseIndexed(cUnit, r0, r12, rLR, 2, kWord);
- case 1: // Get the target compiled code address [use r0, set rLR]
- loadBaseDisp(cUnit, mir, r0, OFFSETOF_MEMBER(Method, compiledInsns),
- rLR, kWord, INVALID_SREG);
+ break;
+ case 1: // Get method->declaring_class_->super_class [usr rLR, set rLR]
+ loadWordDisp(cUnit, rLR, Class::SuperClassOffset().Int32Value(),
+ rLR);
+ break;
+ case 2: // Get ...->super_class_->vtable [u/s rLR]
+ loadWordDisp(cUnit, rLR, Class::VTableOffset().Int32Value(), rLR);
+ break;
+ case 3: // Get target method [use rLR, set r0]
+ loadWordDisp(cUnit, rLR, (target_idx * 4) +
+ art::Array::DataOffset().Int32Value(), r0);
+ break;
+ case 4: // Get the target compiled code address [uses r0, sets rLR]
+ loadWordDisp(cUnit, r0, Method::GetCodeOffset().Int32Value(), rLR);
+ break;
default:
return -1;
}
-#endif
+ return state + 1;
+}
+
+/* Slow-path version of nextSuperCallInsn */
+static int nextSuperCallInsnSP(CompilationUnit* cUnit, MIR* mir,
+ DecodedInstruction* dInsn, int state,
+ ArmLIR* rollback)
+{
+ DCHECK(rollback != NULL);
+ RegLocation rlArg;
+ ArmLIR* skipBranch;
+ ArmLIR* skipTarget;
+ int tReg;
+ /*
+ * This handles the case in which the base method is not fully
+ * resolved at compile time. We must generate code to test
+ * for resolution a run time, bail to the slow path if not to
+ * fill in all the tables. In the latter case, we'll restart at
+ * at the beginning of the sequence.
+ */
+ switch(state) {
+ case 0: // Get the current Method* [sets r0]
+ loadCurrMethodDirect(cUnit, r0);
+ break;
+ case 1: // Get method->dex_cache_resolved_methods_ [usr r0, set rLR]
+ loadWordDisp(cUnit, r0,
+ Method::GetDexCacheResolvedMethodsOffset().Int32Value(), rLR);
+ break;
+ case 2: // method->dex_cache_resolved_methods_->Get(meth_idx) [u/s rLR]
+ loadWordDisp(cUnit, rLR, (dInsn->vB * 4) +
+ art::Array::DataOffset().Int32Value(), rLR);
+ break;
+ case 3: // Resolved?
+ skipBranch = genCmpImmBranch(cUnit, kArmCondNe, rLR, 0);
+ // Slowest path, bail to helper, rollback and retry
+ loadWordDisp(cUnit, rSELF,
+ OFFSETOF_MEMBER(Thread, pResolveMethodFromCode), rLR);
+ loadConstant(cUnit, r1, dInsn->vB);
+ newLIR1(cUnit, kThumbBlxR, rLR);
+ genUnconditionalBranch(cUnit, rollback);
+ // Resume normal slow path
+ skipTarget = newLIR0(cUnit, kArmPseudoTargetLabel);
+ skipTarget->defMask = ENCODE_ALL;
+ skipBranch->generic.target = (LIR*)skipTarget;
+ // Get base_method->method_index [usr rLR, set rLR]
+ loadBaseDisp(cUnit, mir, rLR,
+ Method::GetMethodIndexOffset().Int32Value(), rLR,
+ kUnsignedHalf, INVALID_SREG);
+ // Load "this" [set r1]
+ rlArg = oatGetSrc(cUnit, mir, 0);
+ loadValueDirectFixed(cUnit, rlArg, r1);
+ // Load curMethod->declaring_class_ [uses r0, sets r0]
+ loadWordDisp(cUnit, r0, Method::DeclaringClassOffset().Int32Value(),
+ r0);
+ case 4: // Get method->declaring_class_->super_class [usr r0, set r0]
+ loadWordDisp(cUnit, r0, Class::SuperClassOffset().Int32Value(), r0);
+ break;
+ case 5: // Get ...->super_class_->vtable [u/s r0]
+ loadWordDisp(cUnit, r0, Class::VTableOffset().Int32Value(), r0);
+ // In load shadow fold vtable_ object header size into method_index_
+ opRegImm(cUnit, kOpAdd, rLR,
+ art::Array::DataOffset().Int32Value() / 4);
+ if (!(mir->OptimizationFlags & MIR_IGNORE_RANGE_CHECK)) {
+ // Range check, throw NSM on failure
+ tReg = oatAllocTemp(cUnit);
+ loadWordDisp(cUnit, r0, art::Array::LengthOffset().Int32Value(),
+ tReg);
+ genBoundsCheck(cUnit, tReg, rLR, mir->offset, NULL);
+ oatFreeTemp(cUnit, tReg);
+ }
+ // Get target Method*
+ loadBaseIndexed(cUnit, r0, r0, rLR, 2, kWord);
+ break;
+ case 6: // Get the target compiled code address [uses r0, sets rLR]
+ loadWordDisp(cUnit, r0, Method::GetCodeOffset().Int32Value(), rLR);
+ break;
+ default:
+ return -1;
+ }
return state + 1;
}
@@ -822,6 +868,10 @@
newLIR1(cUnit, kThumbBlxR, rLR);
}
+/*
+ * All invoke-interface calls bounce off of art_invoke_interface_trampoline,
+ * which will locate the target and continue on via a tail call.
+ */
static void genInvokeInterface(CompilationUnit* cUnit, MIR* mir)
{
DecodedInstruction* dInsn = &mir->dalvikInsn;
@@ -847,13 +897,41 @@
DecodedInstruction* dInsn = &mir->dalvikInsn;
int callState = 0;
ArmLIR* nullCk;
+ ArmLIR* rollback;
+ Method* baseMethod = cUnit->method->GetDexCacheResolvedMethods()->
+ Get(dInsn->vB);
+ NextCallInsn nextCallInsn;
+ bool fastPath = true;
+ if (baseMethod == NULL) {
+ fastPath = false;
+ } else {
+ Class* superClass = cUnit->method->GetDeclaringClass()->GetSuperClass();
+ if (superClass == NULL) {
+ fastPath = false;
+ } else {
+ int32_t target_idx = baseMethod->GetMethodIndex();
+ if (superClass->GetVTable()->GetLength() <= target_idx) {
+ fastPath = false;
+ } else {
+ fastPath = (superClass->GetVTable()->Get(target_idx) != NULL);
+ }
+ }
+ }
+ if (fastPath) {
+ nextCallInsn = nextSuperCallInsn;
+ rollback = NULL;
+ } else {
+ nextCallInsn = nextSuperCallInsnSP;
+ rollback = newLIR0(cUnit, kArmPseudoTargetLabel);
+ rollback->defMask = -1;
+ }
// TODO - redundantly loading arg0/r1 ("this")
if (mir->dalvikInsn.opcode == OP_INVOKE_SUPER)
callState = genDalvikArgsNoRange(cUnit, mir, dInsn, callState, &nullCk,
- false, nextSuperCallInsn, NULL);
+ false, nextCallInsn, rollback);
else
callState = genDalvikArgsRange(cUnit, mir, dInsn, callState, &nullCk,
- nextSuperCallInsn, NULL);
+ nextCallInsn, rollback);
// Finish up any of the call sequence not interleaved in arg loading
while (callState >= 0) {
callState = nextSuperCallInsn(cUnit, mir, dInsn, callState, NULL);